On this page · 12 sections
- What changed in US AI export controls, May 2025 to January 2026
- The Nvidia H20 case: why silicon, not software, sets the boundary
- How export controls translate into model choice
- The EU AI Act: GPAI duties since 2 August 2025
- Comparing the three regulatory blocs an enterprise must plan for
- India-specific considerations
- Build, buy, or self-host: the trade-off under two regulators
- Enforcement reality and what it means operationally
- A procurement checklist for compliant model choice in 2026
- FAQ
- How eCorpIT can help
- References
Summary. Between 15 May 2025 and 22 January 2026, US AI chip export policy was rewritten twice: the Commerce Department's Bureau of Industry and Security rescinded the Biden-era AI Diffusion Rule in May 2025, then codified a new framework for advanced AI chips by January 2026. In parallel, the EU AI Act's obligations for general-purpose AI (GPAI) models took effect on 2 August 2025, with penalties reaching up to €35 million or 7% of global annual turnover. Nvidia, whose China-bound H20 chip was modified and then disrupted across May to August 2025, sits at the center of this. For CTOs, the practical question is no longer "which model is best" but "which model can my organisation legally run, in which region, on which silicon, for the next 3 years." This guide covers the rules, the model-choice trade-offs, India's DPDP context, and a procurement checklist.
The cost of getting this wrong is not abstract. A model or a GPU that is compliant in Gurugram may be restricted in Frankfurt and unavailable in Shenzhen. The real risk is usually the supply chain and the data path, not the model weights.
What changed in US AI export controls, May 2025 to January 2026
The Biden administration's AI Diffusion Rule, published in January 2025, would have tiered the world into categories and capped compute exports per country. It never took full effect. On 15 May 2025, the Bureau of Industry and Security (BIS) within the US Department of Commerce announced it would rescind the rule before the compliance date and replace it with alternative guidance.
Legal analysts tracked the shift closely. WilmerHale described the change in a 15 May 2025 client alert, noting that while the formal rule was paused, new guidance raised the compliance risk for AI-related exports rather than lowering it. A&O Shearman and Wiley published parallel analyses on 27 May and 22 May 2025 covering the rescission and the replacement guidance for advanced computing integrated circuits.
By early 2026, the policy had hardened again. Mayer Brown reported on 22 January 2026 that the administration's policies on advanced AI chips had been codified, with effects across the AI ecosystem. Morrison Foerster followed on 9 February 2026 with practical guidance on managing export control risk in the AI chip ecosystem. The Congressional Research Service maintains a standing analysis of US export controls and advanced semiconductors for context on how Washington has used the Export Administration Regulations against China since 2022.
Two practical points follow for enterprise buyers. First, the rules move faster than a typical 3-year hardware refresh cycle, so any plan that assumes today's list of permitted destinations is permanent is fragile. Second, the controls now reach further down the stack than the chip itself, touching cloud access, model weights above certain training thresholds, and the guidance that vendors must follow.
Why the codification matters more than the rescission
It is tempting to read the May 2025 rescission as deregulation. It was not. The Diffusion Rule was replaced, not removed, and the January 2026 codification gave the replacement framework durable legal form. Morrison Foerster's 9 February 2026 guidance on managing export control risk in the AI chip ecosystem treats the controls as a live, enforceable regime that buyers must screen against, not a relic.
For an enterprise, three controls reach the model layer. The first is the chip itself: advanced GPUs above defined performance thresholds need a license to reach certain destinations. The second is cloud compute: providing remote access to controlled compute can itself be a controlled transaction. The third is the model weights: training runs above certain compute thresholds, and the resulting weights, can fall inside the regime. A company that assumes "we only call an API" has no exposure is reading the surface, not the rules.
The Nvidia H20 case: why silicon, not software, sets the boundary
The clearest illustration of how export policy reaches enterprise model choice is the Nvidia H20, a chip designed specifically to stay inside earlier US limits for China. In May 2025, Reuters reported that Nvidia was modifying the H20 for China to stay inside tightening US rules. By 22 August 2025, the picture reversed: reports said Nvidia was looking to halt H20 production after China cracked down on domestic purchases of the chip.
The lesson for a CTO is direct. The same model family can run on different accelerators, but where you can buy and deploy those accelerators is set by export law, not by the model card. If your inference capacity depends on a GPU that becomes restricted, your model availability is hostage to a policy you do not control.
How export controls translate into model choice
Most enterprises do not train frontier models. They consume them, through an API or a self-hosted open-weights checkpoint, on cloud or on-premises GPUs. Export controls touch all four of those decisions. The table below maps the deployment patterns against the regulatory exposure as of February 2026.
| Deployment pattern | Primary regulatory exposure | Practical constraint for enterprises |
|---|---|---|
| Frontier API (US-hosted) | US export rules on cloud compute and model access; EU AI Act GPAI duties on the provider | Region availability and data-residency terms can change; provider must supply EU technical documentation |
| Open-weights, self-hosted | Hardware export controls on the GPUs you buy; AI Act duties may pass to your deployer role | GPU supply risk; you may inherit transparency and copyright obligations |
| Sovereign or regional cloud | Local data law (EU AI Act, India DPDP) plus the cloud's silicon supply | Reduces data-transfer risk but not chip-supply risk |
| China-region deployment | US controls on exporting advanced chips and weights into China | Restricted access to US frontier hardware; reliance on domestic accelerators |
| Edge or air-gapped on-prem | GPU import controls; lighter data-transfer exposure | Hardest to upgrade if your accelerator line is later restricted |
A second decision axis is whether a model counts as general-purpose AI under EU rules, because that determines who carries the compliance load.
The EU AI Act: GPAI duties since 2 August 2025
The EU AI Act applies in phases. Its prohibitions on unacceptable-risk uses began on 2 February 2025. The obligations for providers of general-purpose AI models, the category that covers large foundation models, applied from 2 August 2025, as set out in the European Commission's regulatory framework for AI and tracked on the AI Act implementation timeline. High-risk system rules phase in through 2 August 2026 and beyond, a sequence laid out in the EU AI Act timeline of compliance dates.
The financial stakes are concrete. The Act sets a top penalty tier of up to €35 million or 7% of worldwide annual turnover for the most serious breaches, with lower tiers for other violations. For an enterprise, the duty depends on your role. A provider that places a GPAI model on the EU market carries transparency, documentation and copyright-policy duties. A deployer that integrates a third-party model into a high-risk use case carries its own set of obligations from August 2026.
For model choice, this means a US-hosted frontier API is only usable in the EU if its provider supplies the technical documentation and training-data summaries that EU deployers need to satisfy their own duties. A model that ships without that paperwork is a compliance gap you inherit. We design applications aligned with EU AI Act requirements, which in practice means recording the provider's documentation and your own deployment context before go-live, not after.
Comparing the three regulatory blocs an enterprise must plan for
| Dimension | United States | European Union | India |
|---|---|---|---|
| Lead instrument | Export Administration Regulations (BIS) | EU AI Act | DPDP Act 2023 plus sectoral rules |
| Main lever on model choice | Chip and compute export controls | Risk tiers and GPAI provider duties | Personal-data processing and cross-border transfer |
| Key 2025-2026 date | Diffusion rule rescinded 15 May 2025; chips codified Jan 2026 | GPAI duties from 2 Aug 2025; high-risk from 2 Aug 2026 | DPDP rules operationalising through 2025-2026 |
| Top penalty signal | Export-control enforcement, denial orders | Up to €35M or 7% of global turnover | Penalties up to ₹250 crore per the DPDP Act |
| Net effect for buyers | Constrains where you can buy and run GPUs | Constrains documentation and use cases | Constrains data location and transfer paths |
India-specific considerations
For Indian enterprises and Indian delivery teams serving global clients, two pressures meet. On hardware, India is not a target of US controls in the way China is, so access to US frontier accelerators is broadly available through cloud providers, but the same policy volatility that hit the H20 line can reach the chips Indian data centers rely on. Procurement plans should assume the permitted-destination list can change inside a refresh cycle.
On data, the Digital Personal Data Protection Act 2023 (DPDP) governs how personal data is processed and transferred. The DPDP Act provides for significant financial penalties, up to ₹250 crore for certain breaches, which raises the bar for any model deployment that touches Indian personal data. A model API that sends prompts and personal data outside India needs a transfer basis and a record of where inference happens. For teams building for both Europe and India, the practical path is one architecture that satisfies the stricter of the EU AI Act and DPDP on each axis, rather than two parallel stacks.
Indian firms exporting AI-enabled software also sit downstream of US controls when their products embed US-origin models or run on US-controlled silicon. A SaaS product shipped to a restricted destination can pull export law into a software contract that looks purely commercial on its face.
Build, buy, or self-host: the trade-off under two regulators
The model-sourcing decision used to turn on cost and quality. In 2026 it also turns on who carries the regulatory load. Three patterns dominate, and each shifts the burden differently.
Calling a frontier API from a US provider is the fastest path to capability. The provider owns the training compute and the chip supply, so the export-control exposure largely sits with them. The catch is regional: if the provider cannot offer the service in a region, or cannot supply the EU technical documentation a deployer needs, your roadmap stalls on someone else's compliance posture. You also accept the provider's data-handling terms, which matters under both the EU AI Act and DPDP.
Self-hosting open-weights models moves control to you and moves risk with it. You choose the accelerator, so you also own the GPU supply question that the H20 episode made vivid. You may also inherit deployer duties under the EU AI Act, and if you fine-tune and redistribute, you can edge toward provider duties. The upside is portability: open weights can move between accelerators and regions in a way a closed API cannot.
A sovereign or regional cloud sits between the two. It reduces cross-border data transfer, which helps with DPDP and EU data-residency expectations, but it does not remove the chip-supply question, because that cloud still buys silicon inside the same export regime. The honest summary: no pattern removes export-control exposure, it only moves where the exposure lands.
For most enterprises the answer is a mix. Keep a closed API for breadth of capability, qualify one open-weights model for portability, and document the role you play for each. That hedge is what lets you absorb a rule change without a fire drill.
Enforcement reality and what it means operationally
Two regulators, two enforcement styles. US export controls are enforced through licensing, denial orders and penalties administered by BIS, and the Congressional Research Service's analysis of US export controls and advanced semiconductors documents how aggressively the regime has been used since 2022. The practical signal for a buyer is that the list of restricted parties and destinations is dynamic, and screening is continuous, not a one-time check at procurement.
The EU AI Act carries financial penalties up to €35 million or 7% of worldwide annual turnover for the most serious breaches, a ceiling large enough to change board-level behavior. The enforcement burden is documentation-led: regulators expect you to show the technical records, the role classification, and the risk assessment that justified your deployment. A model you cannot document is a model you cannot defend.
Operationally, this collapses into one discipline: keep a living record that ties each production model to its accelerator, its hosting region, its data path, and its role classification. Most organisations already maintain a software bill of materials. The 2026 addition is a compliance bill of materials for AI, refreshed whenever the model, the silicon, or the rules change. We design applications aligned with EU AI Act and DPDP requirements so that record exists from day one rather than being reconstructed under audit pressure. For teams formalising this, our enterprise generative AI strategy guide covers the governance scaffolding in more depth.
A procurement checklist for compliant model choice in 2026
Treat model selection as a regulated supply-chain decision, not a benchmark contest. The order below puts the constraints that can block deployment first.
- Map the silicon. Identify which accelerators your chosen model runs on in production and confirm those GPUs are available in every region you serve, given controls as of February 2026.
- Pin the region and data path. Document where training, fine-tuning and inference physically happen, and which personal data crosses borders.
- Collect provider documentation. For any GPAI model used in the EU, obtain the provider's technical documentation and training-data summary before launch.
- Classify your role and use case. Decide whether you are a provider or deployer under the EU AI Act, and whether the use case is high-risk, which sets your August 2026 duties.
- Check DPDP exposure. For Indian personal data, confirm a lawful transfer basis and a record of processing location.
- Build an exit path. Keep at least one alternative model and one alternative accelerator qualified, because the rules have changed twice in under a year.
The real cost here is usually the migration, not the integration. A model swap forced by an export-control change is cheap if you designed for portability and expensive if you hard-wired one vendor's silicon into your stack.
For deeper background, see our related guides on enterprise generative AI strategy for 2026 and the broader SEO, AEO and GEO landscape that shapes how regulated AI content is discovered.
FAQ
How eCorpIT can help
eCorpIT is a CMMI Level 5 and MSME-certified technology organisation in Gurugram with senior-led engineering teams and partnerships across AWS, Microsoft and Google. We help CTOs and compliance teams map model choice against US export controls, the EU AI Act and India's DPDP Act, then design portable architectures that survive a regulatory change. To plan a compliant AI deployment, contact our team.
References
_Last updated: 20 February 2026._
